AI face swap with lip sync: 8 ways we tried on the White Chicks car scene
An order on our meme-video app asked for two friends in the White Chicks car scene: the driver sings along to A Thousand Miles, the passenger stares at him. Swapping the faces is easy. Keeping the mouths right is not: the passenger must not look like she is talking, and the driver has to sing "I need you" on the beat. Over three days (November 7 to 9, 2026) we ran 18 paid generations with 8 approaches. No single one got both mouths fully right. Every clip below is cut from those runs, one seat at a time and only from the nose down, so you watch the mouth and not the face. Each plays with the original soundtrack.
Models: kling-4.0-video-edit
The test
The source clip is 12.7 seconds and 16:9 in three shots: the passenger (0 to 2.6 s), the driver singing (2.6 to 10.1 s), and the passenger again holding a small dog. Most runs used only the first 5 seconds, which hold both seats. Two face photos went in, one per seat.
How we judged: crop to the mouth, step through at 24 frames a second next to the source, then watch once at normal speed with sound. At 1 or 2 frames a second a talking mouth can look still; we got that wrong twice before cropping.
The passenger reacts. Lips slightly parted, barely moving: surprise, no words.
The driver turns from profile to camera and starts the line. Wide, flat mouth shapes with the upper teeth showing.
Passenger: parted lips, never talking
Higgsfield Genjutsu (motion transfer, 480p, about $0.80 for 5 s) with one prompt for both seats made the passenger talk. Mentioning that the driver sings seemed to pull her mouth along: when we gave her a prompt of her own with no singing in it, her mouth stopped moving.
But it stopped completely. "Lips closed and still" kept the mouth shut for every frame and lost the surprise. Asking for surprise with parted lips brought the talking back. Using a finished clip of ours as the reference did the same. Our read after six Genjutsu runs ($0.95 each, plus one at 720p for $1.70): it can hold a closed mouth still, but once the mouth is open it adds movement of its own.
MiniMax H3 Max Recast (768p, $0.40 plus a token fee, about 2 minutes) kept the faces closest to the photos, down to an earpiece, but still formed an O around 1.5 seconds. Running sync lipsync-2 on the Recast clip with the passenger part of the audio silenced ($0.25) calmed it down, yet it still opens wider than the source mid-shot.
Kling 4.0 Video Edit (720p, $0.94 for 5 s, 127 seconds) is the only run where the passenger matches the source at 12 frames a second: slightly parted, changing slowly, no word shapes. One side effect: it pulled the camera back to show the shoulders and seat.
Opens and closes like speech, with an O shape.
Still, but shut: the surprise is gone.
The expression is back, and so is the talking.
Lighter than Genjutsu, still an O around 1.5 s.
No word shapes, but opens wider than the source.
Closest to the source: parted, slow, no talking.
Driver: singing on the beat
Here the models fail differently. Genjutsu and Recast follow the timing once the driver faces the camera, but keep the mouth open in the profile frames where the source mouth is shut. Running Kling lipsync over the Recast clip ($0.07) was the only run that closed it there.
Kling 4.0 Video Edit gets the timing and the cut right, then sings the wrong shapes: a tall, round, held-open mouth, more "huh? oh!" than "I need you", where the source is wide and flat with the teeth showing.
On time facing the camera; open in profile where the source is shut.
Similar: the front is fine, the profile is open.
The only one that closes the mouth in profile.
Right timing, wrong shapes: a tall round O instead of wide and flat.
The whole driver shot: the piano intro
In the full 7.5-second driver shot the source only turns his head during the piano intro; the singing starts later. Lip-sync models listen to all of the audio, so both lipsync-2 ($0.38) and Kling lipsync ($0.14) opened the mouth to the piano.
Silencing the audio up to the first sung word fixed most of it. sync lipsync-2-pro ($0.63, about 2 minutes) on the Genjutsu clip with a silent intro gave the best-synced driver of the whole test. It was the version we delivered, and it taught us the last lesson: the customer wrote back that the face looked very strange when the lips moved. Lip-sync models redraw the mouth area, and on a face that is not the source face, that shows.
Only head turns until the line starts.
Natural face, looser sync.
Best sync; the redrawn mouth looked strange to the customer.
Moves to the piano and opens too wide.
What failed outright
Genjutsu object swap was blocked by safety checks on both providers we tried (Higgsfield and each::labs; each::labs billed $0). Our guess is that it refuses real photos swapped into a film with known actors, but neither said why. Motion transfer, Recast, Wan and the lip-sync models were never blocked.
Wan 2.2 Animate Replace ($0.31) copied the expression but not the person: wrong hair, blurred features, ghosting in the first shot, and no sound. We did not take it further.
What we would do now
Start with Kling 4.0 Video Edit on a 5-second cut. It is one request for both seats, $0.12 a second on KeelRouter counted on the source plus the output, and it was the only model that left the passenger alone. If the singer’s mouth is wrong, rerun with the mouth shapes spelled out word by word before reaching for a lip-sync model.
If you do add lip sync, run it on the singer only, silence everything before the first sung word, and show the result with sound to someone who knows the faces. Keep the non-singing seat out of it.
Before writing any prompt, work out what each person feels in each shot. Every expensive retry in this test came from a prompt that only said what not to do with the mouth.
Kling 4.0 Video Edit, two seats in one request
720p, $0.12 a second of source plus output. Our 5-second run took 127 seconds.
POST /v1/video/generations
{
"model": "kling-4.0-video-edit",
"video_url": "https://example.com/car-scene-5s.mp4",
"image_urls": [
"https://example.com/passenger.jpg",
"https://example.com/driver.jpg"
],
"prompt": "Shot 1 (0 to 2.6 s) is the passenger, shot 2 is the driver singing. Replace the passenger with the person in <<<image_1>>> and the driver with the person in <<<image_2>>>. Copy every mouth from <<<video_1>>> frame by frame: the passenger never talks, lips slightly parted in surprise; the driver's lips open and close exactly when the source driver's do, with the same shapes. Keep the head turns, framing, lighting and the cut."
}FAQ
- Which AI model keeps lip sync when swapping faces in a video?
- In our test Kling 4.0 Video Edit came closest in one pass: the non-singing person stayed silent and the singer moved on the beat, though with the wrong mouth shapes. Motion-transfer and recast models made the silent person look like they were talking.
- Should I run a lip-sync model after the face swap?
- Only on the person who sings, with silence before the first sung word. It gave the best sync in our test, but it redraws the mouth, and our customer said the face looked strange when the lips moved.
- How much did the whole test cost?
- About $15 for 18 paid runs, from $0.07 for a lip-sync pass to $1.70 for one 720p Genjutsu run. Most single tries were under $1. Test on 5 seconds at the lowest resolution before paying for the full clip.
More guides
KeelRouter is an independent service and is not affiliated with Kuaishou, Higgsfield, MiniMax, Sync Labs or Alibaba. Model names are trademarks of their owners, used here only to name the model a request runs on.